fix(sdk): use sys.orig_argv for process.command to handle python -m invocations (#5083)

* fix(sdk): use sys.orig_argv for process.command to handle python -m invocations

ProcessResourceDetector populated process.command, process.command_line,
and process.command_args from sys.argv. For applications launched via
`python -m <module>`, the interpreter rewrites sys.argv[0] to the
resolved module path, so the ``-m <module>`` portion of the original
invocation is lost and the detector emits misleading telemetry.

Python 3.10+ exposes sys.orig_argv which preserves the original
arguments received by the interpreter. Since the SDK already requires
Python >= 3.10, switch to sys.orig_argv (with a getattr fallback for
safety). This also aligns with the OTel semantic conventions that
reference /proc/<pid>/cmdline for these attributes.

Fixes #4518

Signed-off-by: Ali <alliasgher123@gmail.com>

* nit: drop sys.orig_argv fallback and use PR number in changelog

Two review nits from MikeGoldsmith:

- sys.orig_argv has been available since Python 3.10; the SDK now
  requires 3.10+ so the getattr fallback is dead code. Use sys.orig_argv
  directly and update the comment.

- CHANGELOG entries should reference the PR number, not the issue
  number.

Signed-off-by: Ali <alliasgher123@gmail.com>

* test: remove redundant sys.argv patch from test_process_detector

Signed-off-by: Ali <alliasgher123@gmail.com>

* test: drop superfluous sys.argv mock in python-m test

The detector reads only sys.orig_argv, so patching sys.argv had no
effect on the test's outcome. Keep only the sys.orig_argv patch.

Signed-off-by: Ali <alliasgher123@gmail.com>

---------

Signed-off-by: Ali <alliasgher123@gmail.com>
Co-authored-by: Mike Goldsmith <goldsmith.mike@gmail.com>
Co-authored-by: Riccardo Magliocchetti <riccardo.magliocchetti@gmail.com>

Author: 3 people

CHANGELOG.md

@@ -28,6 +28,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   ([#5076](https://github.com/open-telemetry/opentelemetry-python/pull/5076))
 - `opentelemetry-semantic-conventions`: use `X | Y` union annotation
   ([#5096](https://github.com/open-telemetry/opentelemetry-python/pull/5096))
+- `opentelemetry-sdk`: Fix `ProcessResourceDetector` to use `sys.orig_argv` so that `process.command`, `process.command_line`, and `process.command_args` reflect the original invocation for `python -m <module>` runs (where `sys.argv[0]` is rewritten to the module path)
+  ([#5083](https://github.com/open-telemetry/opentelemetry-python/pull/5083))
 - `opentelemetry-sdk`: make resource detector ordering deterministic
   ([#5120](https://github.com/open-telemetry/opentelemetry-python/pull/5120))
 - Add WeaverLiveCheck test util

opentelemetry-sdk/src/opentelemetry/sdk/resources/__init__.py

@@ -385,9 +385,16 @@ def detect(self) -> "Resource":
         _process_pid = os.getpid()
         _process_executable_name = sys.executable
         _process_executable_path = os.path.dirname(_process_executable_name)
-        _process_command = sys.argv[0]
-        _process_command_line = " ".join(sys.argv)
-        _process_command_args = sys.argv
+        # Use sys.orig_argv, which preserves the original arguments received
+        # by the interpreter. This correctly captures ``python -m <module>``
+        # invocations where sys.argv is rewritten to the resolved module path
+        # and the ``-m <module>`` information is lost. sys.orig_argv also
+        # aligns with /proc/<pid>/cmdline, which the OTel semantic
+        # conventions reference for these attributes.
+        _process_argv = list(sys.orig_argv)
+        _process_command = _process_argv[0] if _process_argv else ""
+        _process_command_line = " ".join(_process_argv)
+        _process_command_args = _process_argv
         resource_info = {
             PROCESS_RUNTIME_DESCRIPTION: sys.version,
             PROCESS_RUNTIME_NAME: sys.implementation.name,

opentelemetry-sdk/tests/resources/test_resources.py

@@ -574,7 +574,7 @@ def test_service_name_env_precedence(self):
         )
 
     @patch(
-        "sys.argv",
+        "sys.orig_argv",
         ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"],
     )
     def test_process_detector(self):
@@ -620,15 +620,39 @@ def test_process_detector(self):
             os.path.dirname(sys.executable),
         )
         self.assertEqual(
-            aggregated_resource.attributes[PROCESS_COMMAND], sys.argv[0]
+            aggregated_resource.attributes[PROCESS_COMMAND], sys.orig_argv[0]
         )
         self.assertEqual(
             aggregated_resource.attributes[PROCESS_COMMAND_LINE],
-            " ".join(sys.argv),
+            " ".join(sys.orig_argv),
         )
         self.assertEqual(
             aggregated_resource.attributes[PROCESS_COMMAND_ARGS],
-            tuple(sys.argv),
+            tuple(sys.orig_argv),
+        )
+
+    @patch("sys.orig_argv", ["/usr/bin/python", "-m", "myapp"])
+    def test_process_detector_uses_orig_argv_for_python_m(self):
+        """For ``python -m <module>`` invocations sys.argv[0] is rewritten to
+        the resolved module path, losing the ``-m <module>`` information.
+        sys.orig_argv preserves the original invocation and must be preferred.
+        See https://github.com/open-telemetry/opentelemetry-python/issues/4518.
+        """
+        aggregated_resource = get_aggregated_resources(
+            [ProcessResourceDetector()], Resource({"foo": "bar"})
+        )
+
+        self.assertEqual(
+            aggregated_resource.attributes[PROCESS_COMMAND],
+            "/usr/bin/python",
+        )
+        self.assertEqual(
+            aggregated_resource.attributes[PROCESS_COMMAND_LINE],
+            "/usr/bin/python -m myapp",
+        )
+        self.assertEqual(
+            aggregated_resource.attributes[PROCESS_COMMAND_ARGS],
+            ("/usr/bin/python", "-m", "myapp"),
         )
 
     def test_resource_detector_entry_points_default(self):